比赛已经告一段落,现在我们队兑现承诺,将比赛方案开源给大家,互勉互助,共同进步。
队伍介绍
我们的队伍名是Baseline,我们因分享Baseline结缘,所以就把队伍名叫Baseline。
队长: 方曦来自上海交通大学,研三。
队员 :吕晓欣来自网易,AI工程师
队员:王浩来自北京星河亮点,软件研发
队员:杨新达来自广州一家企业,AI工程师
方案
摘要
对于当前通信系统来说,物理层是通信服务得到保障的基础;而对于物理层来说,MIMO则是基本的支撑技术;对于MIMO来说,准确地确定信道质量并做有效反馈及利用又是必不可少的关键问题。
在国际标准化组织3GPP的讨论内,目前这部分工作是通过CSI 参考信号设计及CSI反馈机制完成。在当前的CSI反馈设计中,主要是依赖矢量量化、码本设计的方式来实现信道特征的提取与反馈,例如基于TYPE1、TYPE2的CSI反馈设计等。在目前的实践来看,这类反馈方式是有效的,但是由于其核心思想是基于信息抽取、码本反馈的方式,其所反馈的目标信息实际上是有损信道信息。
在本次大赛中,我们从计算机视觉角度建模,设计出一种基于CNN的自编码器结构。我们采用了带有SE结构的BCSP模块作为网络的基础组件,在计算效率和网络精度上都有较好的效果;采用带有误差恢复能力的量化模块,一方面能降低量化误差,同时也能提高编码器的训练效果;通过分析大赛数据,我们利用Fast-AutoAugment思路找到了4种数据增强方法,完美的解决了在384 附近bit数的网络过拟合问题;我们利用剪枝和降低量化精度的方式,大幅度加速了我们的训练过程。最终我们获得了第7名的好成绩。
关键词
无线通信, 信道反馈, 卷积神经网络, 注意力机制, 数据增强
1 注意力机制的使用
在我们采用的注意力机制是SE-Net: Squeeze-and-Excitation Networks,简称SE-Net[1],它赢得了最后一届ImageNet2017竞赛分类任务的冠军,其基本原理是对于每个输出channel,预测一个常数权重,对每个channel加权一下。结构如下图:
图1:SE 注意力机制
第一步每个通道H*W个数全局平均池化得到一个标量,称之为Squeeze,然后两个FC得到01之间的一个权重值,对原始的每个HxW的每个元素乘以对应通道的权重,得到新的feature map,称之为Excitation。任意的原始网络结构,都可以通过这个Squeeze-Excitation的方式进行feature recalibration,如下图。
图2:SENet基础结构
具体实现上就是一个Global Average Pooling-FC-ReLU-FC-Sigmoid,第一层的FC会把通道降下来,然后第二层FC再把通道升上去,得到和通道数相同的C个权重,每个权重用于给对应的一个通道进行加权。上图中的r就是缩减系数,实验确定选取16,可以得到较好的性能并且计算量相对较小。SENet的核心思想在于通过网络根据loss去学习特征权重,使得有效的feature map权重大,无效或效果小的feature map权重小的方式训练模型达到更好的结果。
我们将SENet一个子结构,嵌入到C3和BottleneckCSP模块的最后一层。如图3所示。
在该赛题中SE结构或者说注意力机制能够大幅度提升模型的拟合能力,让我们的模型能够成功的完成432bit 达标,但随之而来的模型过拟合现象困扰了我们很长时间。
2 量化误差恢复模块
在量化编码过程中,经过量化-反量化操作,将会让原始编码丢失一部分信息,即量化误差,量化误差的存在不仅使得模型最终NMSE会比无量化操作的更高,还会减慢decoder的训练速度和效果。故我们提出量化误差恢复模块,即对反量化后的编码进行refine,使之更加接近无量化损失。
图3:基础模块图
具体操作是,我们对量化-反量化后的编码,通过两层全连接(带有bn和非线性层),并经过标准化处理,得到值域与量化误差值域相同的输出(通过sigmoid以及scale等操作调整值域为[-12B+1,12B+1])以残差的方式加到原始反量化编码之后,以起到恢复量化误差的效果。同时,为了使得这个模块能更好地按设想工作,我们对此模块的输出增加了一路损失函数,使得恢复后的编码与量化前的编码更加接近。
设量化前编码为X,量化后编码为X’,我们的误差恢复模块为R,则额外监督表示为如下:L(X+R(X'), X)。
图4:误差恢复模块与Simsiam结构对比图
上图同对比了误差恢复模块和自监督算法中的SimSiam[2]结构对比图。如果我们把量化误差看座位一种数据增强,那么量化误差修复这一块刚好可以看作一种自监督学习网络,同时存在一条支路在反向传播过程中能够计算出准确的梯度,让我们能够获得更好的encoder层。
3 数据增强
赛方提供的数据200*3000是按序摆放的,通过分析,我们发现3000这个维度中各个数据似乎存在一些相似关系,在我们全部的数据增强过程中都不会去破坏这种模式。数据增强固然可以一定程度上缓解模型过拟合,但是如果设计的不得当,网络会学到很多没用的信息从而不能训到很低的nmse,为此我们借鉴了Fast-AutoAugment中的思想,对于每一种数据增强,我们利用原始数据训练的模型在验证集数据+该数据增强统计nmse,如果nmse过高,那么这种数据增强大概率改变了原始数据分布,不应该背采纳。通过这种方式我们选取了4中数据增强方法:
1-X
实部虚部shuffle
MixUp
CutMix
传统的MixUp和CutMix会破坏数据原有的模式,所以我们对其进行一些改造。在样本采样过程中,我们只会选择属于同一种patten的两个样本进行融合;我们不会去破坏16这个维度的数值关系,所以CutMix过程中随机选择24行中的一部分进行替换,这是因为24这个维度虽然有patten,但是似乎不存在特别明显的数值关系。通过这种方式我们能偶成功的训练出384bit的模型。
图5:数据增强效果图
5 剪枝与量化
量化层我们选择了简单的均匀量化操作。量化bit数目选取上,考虑到任务更加侧重更小传输bit而不是极致的精度(低NMSE),故可以选择使用更小的量化bit数目,而太小的量化bit数目会导致量化误差过大,使得decoder训练更加困难也更容易过拟合。权衡上述,我们选择了使用Bit=3的量化操作。
训练初始模型时,我们首先选择使用bit数为432的bitstream构建模型进行训练,训练完成之后对encoder最末层全连接和decoder最前层全连接进行裁剪,得到384bit的autoencoder模型,然后进行进一步finetune,得到384bit模型(3bit*128)。在比赛的最后阶段,我们选择对128个code中的6个,量化bit从3bit压缩到2bit量化,进一步finetune,得到最终提交的378bit模型。即最终提交的378bit模型中,有122个code采用3bit量化编码,6个code采用2bit量化。
致谢
感谢主办方提供数据,感谢DataFountain平台提供支持和及时的问题反馈!
Code
modelDesign.py
# ======================================================================================================================= # ======================================================================================================================= import numpy as np import torch.nn as nn import torch import torch.nn.functional as F from torch.utils.data import Dataset from collections import OrderedDict NUM_FEEDBACK_BITS_STARTS = 768 NUM_FEEDBACK_BITS = 384 # pytorch版本一定要有这个参数 channel_last = 1 CR_dim = 128 REFINEMENT = 1 class Mish(torch.nn.Module): def __init__(self): super().__init__() def forward(self, x): x = x * (torch.tanh(torch.nn.functional.softplus(x))) return x ACT = nn.SiLU() # ======================================================================================================================= # ======================================================================================================================= # Number to Bit Defining Function Defining def Num2Bit(Num, B): Num_ = Num.type(torch.uint8) def integer2bit(integer, num_bits=B * 2): dtype = integer.type() exponent_bits = -torch.arange(-(num_bits - 1), 1).type(dtype) exponent_bits = exponent_bits.repeat(integer.shape + (1,)) out = integer.unsqueeze(-1) // 2 ** exponent_bits return (out - (out % 1)) % 2 bit = integer2bit(Num_) bit = (bit[:, :, B:]).reshape(-1, Num_.shape[1] * B) return bit.type(torch.float32) def Bit2Num(Bit, B): Bit_ = Bit.type(torch.float32) Bit_ = torch.reshape(Bit_, [-1, int(Bit_.shape[1] / B), B]) num = torch.zeros(Bit_[:, :, 1].shape).cuda() for i in range(B): num = num + Bit_[:, :, i] * 2 ** (B - 1 - i) return num # ======================================================================================================================= # ======================================================================================================================= # Quantization and Dequantization Layers Defining class Quantization(torch.autograd.Function): @staticmethod def forward(ctx, x, B): ctx.constant = B step = 2 ** B out = torch.round(x * step - 0.5) out = Num2Bit(out, B) return out @staticmethod def backward(ctx, grad_output): # return as many input gradients as there were arguments. # Gradients of constant arguments to forward must be None. # Gradient of a number is the sum of its B bits. b, _ = grad_output.shape grad_num = torch.sum(grad_output.reshape(b, -1, ctx.constant), dim=2) / ctx.constant return grad_num, None class Dequantization(torch.autograd.Function): @staticmethod def forward(ctx, x, B): ctx.constant = B step = 2 ** B out = Bit2Num(x, B) out = (out + 0.5) / step return out @staticmethod def backward(ctx, grad_output): # return as many input gradients as there were arguments. # Gradients of non-Tensor arguments to forward must be None. # repeat the gradient of a Num for B time. b, c = grad_output.shape grad_output = grad_output.unsqueeze(2) / ctx.constant grad_bit = grad_output.expand(b, c, ctx.constant) return torch.reshape(grad_bit, (-1, c * ctx.constant)), None class QuantizationLayer(nn.Module): def __init__(self, B): super(QuantizationLayer, self).__init__() self.B = B def forward(self, x): out = Quantization.apply(x, self.B) return out class DequantizationLayer(nn.Module): def __init__(self, B): super(DequantizationLayer, self).__init__() self.B = B def forward(self, x): out = Dequantization.apply(x, self.B) return out # ======================================================================================================================= # ======================================================================================================================= # Encoder and Decoder Class Defining def autopad(k, p=None): # kernel, padding # Pad to 'same' if p is None: p = k // 2 if isinstance(k, int) else [x // 2 for x in k] # auto-pad return p class SEBlock(nn.Module): def __init__(self, input_channels, internal_neurons): super(SEBlock, self).__init__() self.down = nn.Conv2d(in_channels=input_channels, out_channels=internal_neurons, kernel_size=1, stride=1, bias=True, padding_mode='circular') self.up = nn.Conv2d(in_channels=internal_neurons, out_channels=input_channels, kernel_size=1, stride=1, bias=True, padding_mode='circular') def forward(self, inputs): x = F.avg_pool2d(inputs, kernel_size=inputs.size(3)) x = self.down(x) x = F.leaky_relu(x) x = self.up(x) x = torch.sigmoid(x) x = x.repeat(1, 1, inputs.size(2), inputs.size(3)) return inputs * x class Conv(nn.Module): # Standard convolution def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True): # ch_in, ch_out, kernel, stride, padding, groups super(Conv, self).__init__() self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g, bias=False) self.bn = nn.BatchNorm2d(c2) self.act = ACT def forward(self, x): return self.act(self.bn(self.conv(x))) class Bottleneck(nn.Module): # Standard bottleneck def __init__(self, c1, c2, shortcut=True, g=1, e=0.5): # ch_in, ch_out, shortcut, groups, expansion super(Bottleneck, self).__init__() c_ = int(c2 * e) # hidden channels self.cv1 = Conv(c1, c_, 1, 1) self.cv2 = Conv(c_, c2, 3, 1, g=g) self.add = shortcut and c1 == c2 def forward(self, x): return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x)) class BottleneckCSP(nn.Module): # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion super(BottleneckCSP, self).__init__() c_ = int(c2 * e) # hidden channels self.cv1 = Conv(c1, c_, 1, 1) self.cv2 = nn.Conv2d(c1, c_, 1, 1, bias=False) self.cv3 = nn.Conv2d(c_, c_, 1, 1, bias=False) self.cv4 = Conv(2 * c_, c2, 1, 1) self.bn = nn.BatchNorm2d(2 * c_) # applied to cat(cv2, cv3) self.act = nn.LeakyReLU(0.1, inplace=True) self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)]) self.att = SEBlock(c2, c2 // 2) def forward(self, x): y1 = self.cv3(self.m(self.cv1(x))) y2 = self.cv2(x) return self.att(self.cv4(self.act(self.bn(torch.cat((y1, y2), dim=1))))) class C3(nn.Module): # CSP Bottleneck with 3 convolutions def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion super(C3, self).__init__() c_ = int(c2 * e) # hidden channels self.cv1 = Conv(c1, c_, 1, 1) self.cv2 = Conv(c1, c_, 1, 1) self.cv3 = Conv(2 * c_, c2, 1) # act=FReLU(c2) self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)]) # self.m = nn.Sequential(*[CrossConv(c_, c_, 3, 1, g, 1.0, shortcut) for _ in range(n)]) self.att = SEBlock(c2, c2 // 2) def forward(self, x): return self.att(self.cv3(torch.cat((self.m(self.cv1(x)), self.cv2(x)), dim=1))) class Focus(nn.Module): # Focus wh information into c-space def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True): # ch_in, ch_out, kernel, stride, padding, groups super(Focus, self).__init__() self.conv = Conv(c1 * 4, c2, k, s, p, g, act) # self.contract = Contract(gain=2) def forward(self, x): # x(b,c,w,h) -> y(b,4c,w/2,h/2) return self.conv(torch.cat([x[..., ::2, ::2], x[..., 1::2, ::2], x[..., ::2, 1::2], x[..., 1::2, 1::2]], 1)) # return self.conv(self.contract(x)) class Contract(nn.Module): # Contract width-height into channels, i.e. x(1,64,80,80) to x(1,256,40,40) def __init__(self, gain=2): super().__init__() self.gain = gain def forward(self, x): N, C, H, W = x.size() # assert (H / s == 0) and (W / s == 0), 'Indivisible gain' s = self.gain x = x.view(N, C, H // s, s, W // s, s) # x(1,64,40,2,40,2) x = x.permute(0, 3, 5, 1, 2, 4).contiguous() # x(1,2,2,64,40,40) return x.view(N, C * s * s, H // s, W // s) # x(1,256,40,40) class Expand(nn.Module): # Expand channels into width-height, i.e. x(1,64,80,80) to x(1,16,160,160) def __init__(self, c1, c2, gain=2, k=1, s=1, p=None, g=1, act=True): super().__init__() self.gain = gain self.conv = Conv(c1 // 4, c2, k, s, p, g, act) def forward(self, x): N, C, H, W = x.size() # assert C / s ** 2 == 0, 'Indivisible gain' s = self.gain x = x.view(N, s, s, C // s ** 2, H, W) # x(1,2,2,16,80,80) x = x.permute(0, 3, 4, 1, 5, 2).contiguous() # x(1,16,80,2,80,2) return self.conv(x.view(N, C // s ** 2, H * s, W * s)) # x(1,16,160,160) def conv3x3(in_planes, out_planes, stride=1): """3x3 convolution with padding""" return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=True) class WLBlock(nn.Module): def __init__(self, paths, in_c, k=16, n=[1, 1], e=[1.0, 1.0], quantization=True): super(WLBlock, self).__init__() self.paths = paths self.n = n self.e = e self.k = k self.in_c = in_c for i in range(self.paths): self.__setattr__(str(i), nn.Sequential(OrderedDict([ ("Conv0", Conv(self.in_c, self.k, 3)), ("BCSP_1", BottleneckCSP(self.k, self.k, n=self.n[i], e=self.e[i])), ("C3_1", C3(self.k, self.k, n=self.n[i], e=self.n[i])), ("Conv1", Conv(self.k, self.k, 3)), ]))) self.conv1 = conv3x3(self.k * self.paths, self.k) def forward(self, x): outs = [] for i in range(self.paths): _ = self.__getattr__(str(i))(x) outs.append(_) out = torch.cat(tuple(outs), dim=1) out = self.conv1(out) out = out + x if self.in_c == self.k else out return out class Encoder(nn.Module): B = 3 def __init__(self, feedback_bits, quantization=True): super(Encoder, self).__init__() self.feedback_bits = feedback_bits self.k = 256 self.encoder1 = nn.Sequential(OrderedDict([ ("Conv0", Conv(2, 16, 5)), ("BCSP_1", BottleneckCSP(16, 16, n=2, e=0.5)), ("C3_1", C3(16, 16, n=1, e=2.0)), ("Conv1", Conv(16, self.k, 3)) ])) self.encoder2 = nn.Sequential(OrderedDict([ ("Focus0", Focus(2, 16)), ("BCSP_1", BottleneckCSP(16, 16, n=1, e=1.0)), ("C3_1", C3(16, 16, n=2, e=2.0)), ("Expand0", Expand(16, 16)), ("Conv1", Conv(16, self.k, 3)) ])) self.encoder3 = nn.Sequential(OrderedDict([ ("Conv0", Conv(2, 32, 3)), ("WLBlock1", WLBlock(3, 32, 32, [1, 2, 3], [0.5, 1, 1.5])), ("WLBlock2", WLBlock(2, 32, 32, [2, 4], [1, 2])), ("Conv1", Conv(32, self.k, 3)), ])) self.encoder_conv = nn.Sequential(OrderedDict([ ("conv1x1", Conv(self.k * 3, 2, 1)), ])) self.fc = nn.Linear(768, int(NUM_FEEDBACK_BITS_STARTS / self.B)) self.dim_verify = nn.Linear(int(NUM_FEEDBACK_BITS_STARTS / self.B), int(self.feedback_bits / self.B)) self.sig = nn.Sigmoid() self.quantize = QuantizationLayer(self.B) self.quantization = quantization def forward(self, x): if channel_last: x = x.permute(0, 3, 1, 2).contiguous() x0 = x.view(-1, 768) encoder1 = self.encoder1(x) encoder2 = self.encoder2(x) encoder3 = self.encoder3(x) out = torch.cat((encoder1, encoder2, encoder3), dim=1) out = self.encoder_conv(out) out = out.view(-1, 768) + x0 out = self.fc(out) out = self.dim_verify(out) out = self.sig(out) enq_data = out if self.quantization: out = self.quantize(out) elif self.quantization == 'check': out = out else: out = self.fake_quantize(out) return out, enq_data class Decoder(nn.Module): B = 3 def __init__(self, feedback_bits, quantization=True): super(Decoder, self).__init__() self.k = 64 self.feedback_bits = feedback_bits self.dequantize = DequantizationLayer(self.B) self.dim_verify = nn.Linear(int(self.feedback_bits / self.B), int(NUM_FEEDBACK_BITS_STARTS / self.B)) self.fc = nn.Linear(int(NUM_FEEDBACK_BITS_STARTS / self.B), 768) self.ende_refinement = nn.Sequential( nn.Linear(int(self.feedback_bits / self.B), int(self.feedback_bits / self.B)), nn.BatchNorm1d(int(self.feedback_bits / self.B)), nn.ReLU(True), nn.Linear(int(self.feedback_bits / self.B), int(self.feedback_bits / self.B), bias=False), nn.Sigmoid(), ) self.decoder1 = nn.Sequential(OrderedDict([ ("Conv0", Conv(2, 16, 3)), ("BCSP_1", BottleneckCSP(16, 16, n=1, e=1.0)), ("Conv1", Conv(16, self.k, 1)), ])) self.decoder2 = nn.Sequential(OrderedDict([ ("Conv0", Conv(2, 32, 5)), ("BCSP_1", BottleneckCSP(32, 32, n=4, e=2.0)), ("Conv1", Conv(32, self.k, 1)), ])) self.decoder3 = nn.Sequential(OrderedDict([ ("Conv0", Conv(2, 32, (1, 3))), ("BCSP_1", BottleneckCSP(32, 32, n=4, e=2.0)), ("Conv1", Conv(32, self.k, 1)), ])) self.decoder4 = nn.Sequential(OrderedDict([ ("Conv0", Conv(2, 32, (3, 1))), ("BCSP_1", BottleneckCSP(32, 32, n=4, e=2.0)), ("Conv1", Conv(32, self.k, 1)), ])) self.decoder5 = nn.Sequential(OrderedDict([ ("Focus0", Focus(2, self.k)), ("WLBlock1", WLBlock(3, self.k, self.k, [1, 2, 3], [0.5, 1, 1.5])), ("WLBlock2", WLBlock(2, self.k, self.k, [2, 4], [1, 2])), ("Expand0", Expand(self.k, self.k)), ("Conv1", Conv(self.k, self.k, 1)), ])) self.decoder6 = nn.Sequential(OrderedDict([ ("Conv0", Conv(2, 32, (3, 5))), ("BCSP_1", BottleneckCSP(32, 32, n=4, e=2.0)), ("Conv1", Conv(32, self.k, 5)), ])) self.decoder7 = nn.Sequential(OrderedDict([ ("Conv0", Conv(2, 32, (5, 3))), ("BCSP_1", BottleneckCSP(32, 32, n=4, e=2.0)), ("Conv1", Conv(32, self.k, 3)), ])) self.decoder8 = nn.Sequential(OrderedDict([ ("Focus0", Focus(2, self.k, 5)), ("WLBlock1", WLBlock(2, self.k, self.k, [1, 2], [0.5, 1])), ("WLBlock2", WLBlock(2, self.k, self.k, [1, 2], [1, 0.5])), ("Expand0", Expand(self.k, self.k)), ("Conv1", Conv(self.k, self.k, 5)), ])) if REFINEMENT: self.refinemodel = nn.Sequential(OrderedDict([ ("Conv0", Conv(2, 64, 3)), ("WLBlock1", WLBlock(3, 64, 64, [1, 2, 3], [0.5, 1, 1.5])), ("WLBlock2", WLBlock(2, 64, 64, [2, 4], [1, 2])), ("WLBlock3", WLBlock(2, 64, 64, [2, 4], [1, 2])), ("WLBlock4", WLBlock(2, 64, 64, [1, 3], [1, 2])), ("Conv1", Conv(64, 2, 3)), ])) self.decoder_conv = conv3x3(self.k * 8, 2) self.sig = nn.Sigmoid() self.quantization = quantization def forward(self, x): if self.quantization: out = self.dequantize(x) else: out = x out = out.view(-1, int(self.feedback_bits / self.B)) out_error = self.ende_refinement(out) out = out + out_error - 0.5 deq_data = out out = self.dim_verify(out) out = self.sig(self.fc(out)) out = out.view(-1, 2, 24, 16) out0 = out out1 = self.decoder1(out) out2 = self.decoder2(out) out3 = self.decoder3(out) out4 = self.decoder4(out) out5 = self.decoder5(out) out6 = self.decoder6(out) out7 = self.decoder7(out) out8 = self.decoder8(out) out = torch.cat((out1, out2, out3, out4, out5, out6, out7, out8), dim=1) out = self.decoder_conv(out) + out0 out = self.sig(out) if REFINEMENT: out = self.sig(self.refinemodel(out)) - 0.5 + out if channel_last: out = out.permute(0, 2, 3, 1) return out, deq_data class AutoEncoder(nn.Module): def __init__(self, feedback_bits): super(AutoEncoder, self).__init__() self.encoder = Encoder(feedback_bits) self.decoder = Decoder(feedback_bits) def forward(self, x): feature, enq_data = self.encoder(x) out, deq_data = self.decoder(feature) return out, feature, enq_data, deq_data # ======================================================================================================================= # ======================================================================================================================= # NMSE Function Defining def NMSE(x, x_hat): x_real = np.reshape(x[:, :, :, 0], (len(x), -1)) x_imag = np.reshape(x[:, :, :, 1], (len(x), -1)) x_hat_real = np.reshape(x_hat[:, :, :, 0], (len(x_hat), -1)) x_hat_imag = np.reshape(x_hat[:, :, :, 1], (len(x_hat), -1)) x_C = x_real - 0.5 + 1j * (x_imag - 0.5) x_hat_C = x_hat_real - 0.5 + 1j * (x_hat_imag - 0.5) power = np.sum(abs(x_C) ** 2, axis=1) mse = np.sum(abs(x_C - x_hat_C) ** 2, axis=1) nmse = np.mean(mse / power) return nmse def Score(NMSE): score = 1 - NMSE return score def NMSE_cuda(x, x_hat): x_real = x[:, 0, :, :].view(len(x), -1) - 0.5 x_imag = x[:, 1, :, :].view(len(x), -1) - 0.5 x_hat_real = x_hat[:, 0, :, :].view(len(x_hat), -1) - 0.5 x_hat_imag = x_hat[:, 1, :, :].view(len(x_hat), -1) - 0.5 power = torch.sum(x_real ** 2 + x_imag ** 2, axis=1) mse = torch.sum((x_real - x_hat_real) ** 2 + (x_imag - x_hat_imag) ** 2, axis=1) nmse = mse / power return nmse class NMSELoss(nn.Module): def __init__(self, reduction='sum'): super(NMSELoss, self).__init__() self.reduction = reduction def forward(self, x_hat, x): nmse = NMSE_cuda(x, x_hat) if self.reduction == 'mean': nmse = torch.mean(nmse) else: nmse = torch.sum(nmse) return nmse # ======================================================================================================================= # ======================================================================================================================= import random # Data Loader Class Defining class DatasetFolder(Dataset): def __init__(self, matData, phase='val'): self.matdata = matData self.phase = phase def __getitem__(self, index): y = self.matdata[index] if self.phase == 'train' and random.random() < -0.5: y = y[::-1, :, :].copy() if self.phase == 'train' and random.random() < 0.5: y = y[:, ::-1, :].copy() if self.phase == 'train' and random.random() < 0.5: y = 1 - self.matdata[index] # 数据中存在类似正交的关系 if self.phase == 'train' and random.random() < 0.5: _ = y _[:, :, 0] = y[:, :, 1] _[:, :, 1] = y[:, :, 0] y = _ # 不同时刻数据实虚存在部分相等的情况 if self.phase == 'train' and random.random() < 0.5: index_ = random.randint(0, self.matdata.shape[0] // 3000 - 1) * 3000 + index % 3000 p = random.random() rows = max(int(24 * p), 1) _rows = [i for i in range(24)] random.shuffle(_rows) _rows = _rows[:rows] if random.random() < 0.7: y[_rows] = self.matdata[index_][_rows] # 不同采样点按行合并,保持采样点独有特性,减轻模型对24那个维度的依赖 else: y = (1 - p * 0.2) * y + (p * 0.2) * self.matdata[index_] # 增加数值扰动,保持采样点独有特性 return y def __len__(self): return self.matdata.shape[0]
modelTrain.py
#======================================================================================================================= #======================================================================================================================= import numpy as np import torch from modelDesign import AutoEncoder,DatasetFolder,NUM_FEEDBACK_BITS,NUM_FEEDBACK_BITS_STARTS,NMSELoss,channel_last #* import os import torch.nn as nn import scipy.io as sio import random from torch.cuda.amp import autocast, GradScaler def NMSE_cuda1(x, x_hat): x_real = x[:, :, :, 0].view(len(x),-1) - 0.5 x_imag = x[:, :, :, 1].view(len(x),-1) - 0.5 x_hat_real = x_hat[:, :, :, 0].view(len(x_hat), -1) - 0.5 x_hat_imag = x_hat[:, :, :, 1].view(len(x_hat), -1) - 0.5 power = torch.sum(x_real**2 + x_imag**2, axis=1) mse = torch.sum((x_real-x_hat_real)**2 + (x_imag-x_hat_imag)**2, axis=1) nmse = mse/power return nmse class NMSELoss1(nn.Module): def __init__(self, reduction='sum'): super(NMSELoss1, self).__init__() self.reduction = reduction def forward(self, x_hat, x): nmse = NMSE_cuda1(x, x_hat) if self.reduction == 'mean': nmse = torch.mean(nmse) else: nmse = torch.sum(nmse) return nmse #======================================================================================================================= #======================================================================================================================= # Parameters Setting for Data CHANNEL_SHAPE_DIM1 = 24 CHANNEL_SHAPE_DIM2 = 16 CHANNEL_SHAPE_DIM3 = 2 # Parameters Setting for Training BATCH_SIZE = 64 EPOCHS = 1000 LEARNING_RATE = 1e-5 PRINT_RREQ = 100 #NUM_FEEDBACK_BITS =NUM_FEEDBACK_BITS_3 torch.manual_seed(42) random.seed(42) #======================================================================================================================= #======================================================================================================================= def load_pretrained_weights(model,model_path): encoder_pretrained = torch.load(model_path)['state_dict'] model_dict = model.state_dict() #pretrained_weights ={k:v for k,v in encoder_pretrained.items() if (k in model_dict and 'dim_verify' not in k and 'ende_refinement' not in k and 'fc' not in k)} pretrained_weights ={k:v for k,v in encoder_pretrained.items() if (k in model_dict )} # prune dim_verify layer if 0 and NUM_FEEDBACK_BITS != NUM_FEEDBACK_BITS_STARTS: w = encoder_pretrained['dim_verify.weight'] b = encoder_pretrained['dim_verify.bias'] if model_dict['dim_verify.weight'].shape[0] != encoder_pretrained['dim_verify.weight'].shape[0]: dim = -1 bits_num =model_dict['dim_verify.weight'].shape[0] long = encoder_pretrained['dim_verify.weight'].shape[0] else: dim = 0 bits_num =model_dict['dim_verify.weight'].shape[1] long = encoder_pretrained['dim_verify.weight'].shape[1] #importance = abs(w).sum(dim) #sorted_index = torch.argsort(-1*importance) # descend start = (long -bits_num)//2 end = bits_num + (long - bits_num)//2 if dim == -1: pretrained_weights['dim_verify.weight'] = w[start:end,:] else: pretrained_weights['dim_verify.weight'] = w[:,start:end] model_dict.update(pretrained_weights) model.load_state_dict(model_dict) return model # Model Constructing autoencoderModel = AutoEncoder(NUM_FEEDBACK_BITS) # model_path = './modelSubmit/encoder.pth.tar' # autoencoderModel.encoder =load_pretrained_weights(autoencoderModel.encoder,model_path) # model_path = './modelSubmitTeacher/decoder.pth.tar' # autoencoderModel.decoder =load_pretrained_weights(autoencoderModel.decoder,model_path) model_path = './modelSubmit/encoder.pth.tar' autoencoderModel.encoder.load_state_dict(torch.load(model_path)['state_dict']) model_path = './modelSubmit/decoder.pth.tar' autoencoderModel.decoder.load_state_dict(torch.load(model_path)['state_dict']) #======================================================================================================================= #======================================================================================================================= # Data Loading mat = sio.loadmat('channelData/H_4T4R.mat') data = mat['H_4T4R'] data = data.astype('float32') data = np.reshape(data, (-1, CHANNEL_SHAPE_DIM1, CHANNEL_SHAPE_DIM2, CHANNEL_SHAPE_DIM3)) if not channel_last: data = np.transpose(data, (0, 3, 1, 2)) #random.shuffle(data) split = int(data.shape[0] * 0.95) data_train0, data_test = data[:split], data[split:] random.shuffle(data_train0) split = int(data_train0.shape[0]*0.95) data_train, data_val = data_train0[:split],data_train0[split:] train_dataset = DatasetFolder(data_train0,'train') train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=BATCH_SIZE, shuffle=True, num_workers=0, pin_memory=True) val_dataset = DatasetFolder(data_val,'val') val_loader = torch.utils.data.DataLoader(val_dataset, batch_size=BATCH_SIZE, shuffle=False, num_workers=0, pin_memory=True) test_dataset = DatasetFolder(data_test,'val') test_loader = torch.utils.data.DataLoader(test_dataset, batch_size=BATCH_SIZE, shuffle=False, num_workers=0, pin_memory=True) #======================================================================================================================= #======================================================================================================================= #autoencoderModel = autoencoderModel.cuda() autoencoderModel = torch.nn.DataParallel(autoencoderModel.cuda()) ctl = NMSELoss1(reduction='mean') if channel_last else NMSELoss(reduction='mean') criterion = ctl #nn.MSELoss() criterion_test = ctl feature_criterion = nn.MSELoss() optimizer = torch.optim.AdamW(autoencoderModel.parameters(), lr=LEARNING_RATE) scheduler = torch.optim.lr_scheduler.CosineAnnealingWarmRestarts(optimizer, T_0=20, T_mult=1, eta_min=1e-9, last_epoch=-1) #======================================================================================================================= #======================================================================================================================= # Model Training and Saving bestLoss = 0.105 valLoss = 1e-5 for epoch in range(EPOCHS): scaler = GradScaler() print('lr:',optimizer.param_groups[0]['lr']) autoencoderModel.train() for i, autoencoderInput in enumerate(train_loader): autoencoderInput = autoencoderInput.cuda() with autocast(): autoencoderOutput,_, enq, deq = autoencoderModel(autoencoderInput) loss1 = criterion(autoencoderOutput, autoencoderInput) loss2 = feature_criterion(enq, deq) loss = loss1+0*loss2 scaler.scale(loss).backward() scaler.step(optimizer) scaler.update() optimizer.zero_grad() if i % PRINT_RREQ == 0: print('Epoch: [{0}][{1}/{2}]\t' 'Loss {loss:.4f}\t,Loss_nmse {loss_nmse:.4f}\t,Loss ende {loss_q:.4f}\t'.format(epoch, i, len(train_loader), loss=loss.item(),loss_nmse=loss1.item(),loss_q=loss2.item())) # if (i+1) % (4*PRINT_RREQ) == 0: # break # Model Evaluating autoencoderModel.eval() totalLoss = 0 hist =0 with torch.no_grad(): for i, autoencoderInput in enumerate(val_loader): autoencoderInput = autoencoderInput.cuda() autoencoderOutput, feature, enq, deq = autoencoderModel(autoencoderInput) hist = hist+feature.sum(0)/autoencoderInput.shape[0] totalLoss += criterion_test(autoencoderOutput, autoencoderInput).item()*autoencoderInput.shape[0] averageLoss = totalLoss / len(test_dataset) loss2 = feature_criterion(enq, deq) print('==random split test step==') print(np.std(hist.cpu().numpy())) print(averageLoss,loss2.item()) valavgloss = averageLoss totalLoss = 0 hist =0 with torch.no_grad(): for i, autoencoderInput in enumerate(test_loader): autoencoderInput = autoencoderInput.cuda() autoencoderOutput, feature, enq, deq = autoencoderModel(autoencoderInput) hist = hist+feature.sum(0)/autoencoderInput.shape[0] totalLoss += criterion_test(autoencoderOutput, autoencoderInput).item()*autoencoderInput.shape[0] averageLoss = totalLoss / len(test_dataset) loss2 = feature_criterion(enq, deq) print('==last split test step==') print(np.std(hist.cpu().numpy())) print(averageLoss,loss2.item()) if averageLoss < bestLoss: # Model saving # Encoder Saving torch.save({'state_dict': autoencoderModel.module.encoder.state_dict(), }, './modelSubmit/encoder.pth.tar') # Decoder Saving torch.save({'state_dict': autoencoderModel.module.decoder.state_dict(), }, './modelSubmit/decoder.pth.tar') print("Model saved,avgloss:",averageLoss) bestLoss = averageLoss valLoss = valavgloss print('==show best==') print('valloss:', valLoss, 'testloss:',bestLoss) if epoch>0*50: scheduler.step() #break #======================================================================================================================= #=======================================================================================================================